Browse All Artificial Intelligence Content (818)

John Edward explains the planner-executor loop as a practical architecture for building agentic systems on Azure, focusing on how to separate planning from execution, validate model outputs, and run tool calls safely. The article maps the pattern to Azure services and highlights reliability, observability, and access-control decisions needed for production use.

dchelupati explains how to scale Azure SRE Agent from a single team to enterprise-wide use, focusing on access control, guardrails, auditing, multi-agent topologies, and “agent as code” practices for repeatable deployment and governance.

dchelupati shares a practical “start small” adoption guide for Azure SRE Agent, focused on picking a first operational use case, connecting only the required context (telemetry, repos, runbooks), and expanding into repeatable workflows with the right triggers and permissions.

Sandeep Sen announces the stable Azure MCP Server 3.0 release and explains what’s new for bringing Azure operational tooling into MCP-compatible AI agent workflows, including Terraform and Advisor tooling, more predictable client integration, and security-focused defaults for resource creation.

Jakub Oleksy reports on GitHub’s September 2026 reliability work and incidents, including database pressure reductions, overload controls, added headroom for GitHub Actions, expanded failure containment for pull requests, and progress moving production workloads and traffic to Azure.

Allison announces updates to HydraFusion in GitHub Copilot, including a refreshed frontier-model pool, improved Cascade and Critique orchestration workflows, and a new HydraFusion Max option for deeper reasoning on demanding tasks across VS Code, the Copilot app, and Copilot CLI.

paulyuk shows how the Azure SRE Agent canvas fits into GitHub Copilot (App and Copilot CLI) so on-call engineers can validate an agent’s incident diagnosis with their own telemetry, queries, and access controls, then share a reviewed update back to the team.

The Visual Studio Code Team shares early release notes for VS Code 1.143 Insiders, highlighting updates to MCP server configuration, Agent Host chat controls for background shells, and a new extension contribution point for Explorer title actions.

Microsoft Developer demonstrates an end-to-end workflow for building and shipping a hosted AI agent from the terminal using GitHub Copilot CLI and Microsoft Foundry, including planning, resource creation, model selection, evaluation-driven improvements, and a secure Azure Functions proxy to connect the agent to a website chat experience.

Joe Filcik and Colin Murphy explain how Azure Content Understanding uses grounding and field-level confidence to make document extraction results inspectable and actionable, then show how to calibrate confidence thresholds (per document type and per field) to route work in human-in-the-loop review workflows.

James Rempt shares takeaways from VSLive! San Diego on how experienced Visual Studio developers are adopting GitHub Copilot and other AI tools, while spending more time validating generated code and leaning on hands-on learning. The post also highlights how Visual Studio Subscription benefits support ongoing skill-building.

John Savill gives a quick overview of “computer use” in AI: why a model alone isn’t enough, what an agent harness provides, and how desktop-style interaction changes the way AI systems can execute tasks while still needing boundaries, governance, and visibility.

John Edward breaks down how to design stateful AI agents that can keep context across interactions, safely call external tools, and coordinate multi-step work with structured planning. The article focuses on practical system design: what to store, how to retrieve it, and how to control execution so agents behave reliably in production.

Jacquie Capur joins Matty Stratton to unpack what “meaningful control” over tooling looks like in practice: understanding dependencies, verifying reliability, and keeping a realistic exit path. They connect classic DevOps lessons (config management, Terraform, CI/CD) to today’s fast-moving AI tooling, open models, and data residency constraints.

sbaynes outlines how Microsoft is scaling Azure’s AI infrastructure end-to-end, covering custom silicon (Maia and Cobalt), networking for large AI clusters, and datacenter power and cooling redesigns. The post also highlights hardware-rooted security work like Azure Integrated HSM and the open-source Tessera Integrated HSM initiative.

sbaynes breaks down the Azure Maia 300 system architecture and the design choices Microsoft made to optimize cloud-scale LLM inference, focusing on dataflow, memory locality, low-precision formats, and rack-to-superpod networking to improve tokens-per-watt and tokens-per-dollar.

Ali Zaman explains how Microsoft Discovery coordinates AI agents to help scientists reduce a large research space into a smaller set of promising candidates, speeding up scientific discovery and reducing the number of experiments needed.

Allison announces general availability of the Gemini 4 Argon model in GitHub Copilot, including where developers can select it, which Copilot plans get access, and how admins can control availability via model policy and default model enablement settings.

sbaynes explains what changed when an AI agent moved from summarizing production network incidents to proposing (and potentially executing) real fixes, and why “safe autonomy” needs multiple deterministic layers rather than a single guardrail.

Batu Ozturkler explains how GitHub Secret Scanning reduced false positives and sped up scanning by replacing an LLM-based stage with a compact, fine-tuned ModernBERT classifier that uses surrounding code and text context to decide whether a candidate string is a real credential.

mk_sunitha introduces Fabric App functions (preview), a TypeScript-based serverless capability for Fabric Apps that lets builders run secure server-side logic, use application authentication to access downstream resources, and manage secrets safely. The post also shows how to scaffold, debug, and deploy functions, including using GitHub Copilot in VS Code.

mariozeng explains how Microsoft is rethinking power delivery for Azure AI infrastructure, focusing on grid-to-chip efficiency for high-density GPU systems. The post covers LVDC and 800VDC distribution, disaggregated rack power (Mt. Diablo), solid-state transformers with energy storage, and server-to-silicon power delivery design.

sdighe explains how large-scale AI training and inference are pushing datacenter networks to evolve, and why networking now needs to be designed as part of the AI system. The post highlights open specifications like Multipath Reliable Connection (MRC), plus industry efforts around optical interconnects and Ethernet scale-up networking.

Allison announces updates to GitHub Copilot enterprise managed settings enforcement, including a startup policy freshness requirement and minimum client version enforcement for the sandbox managed setting key.

Ali Zaman explains Microsoft Discovery and how it coordinates specialized AI agents to help scientists run an evidence-driven research loop, using battery design as the example and focusing on why search and summarization aren’t enough for real R&D workflows.

Chu Lahlou explains how Azure Content Understanding prebuilt analyzers go beyond a schema by packaging domain knowledge with grounding and confidence signals, so document extraction results are safer to automate. The post also shares benchmark results for Advanced Contextualization in Content Understanding 2.0 Preview, covering quality, modeled cost, and latency.

Allison announces new Security Overview metrics for GitHub code scanning AI Scan on pull requests, helping org and enterprise admins track findings generated and resolved, remediation rate, repo-level trends, and how often developers apply autofix suggestions.

Allison announces new GitHub Copilot enterprise managed telemetry controls that let admins attribute Copilot activity to a specific user while keeping prompts, responses, and tool content disabled by default, with fine-grained OpenTelemetry capture settings across Copilot CLI, the Copilot app, and VS Code Agent Host sessions.

sbaynes introduces StudentSim, a research system that trains personalized AI “student simulators” from de-identified learner records to evaluate and improve AI tutors before running costly human studies. The post explains how the simulators are measured, trained, and used as a feedback signal for tutor optimization across chess, writing, and math.

Christopher Harrison explains how to make GitHub Copilot more effective on a real codebase by storing project context in the repo. He covers repository and path-specific instructions, reusable “agent skills” for repeatable procedures, custom agents for broader work, and MCP servers (including Playwright) to connect Copilot to external tools.

sbaynes announces Microsoft’s plan to scale the Azure Maia accelerator system to multi-gigawatt deployments starting in 2027, outlining how Maia 200 is already running production inference and what architectural changes in Maia 300 target long-context, latency-sensitive, and agentic workloads.

Wuyi Weng shows how to run Qwen3.8-Flash-Next across two H100 GPUs on an Azure Spot VM, then instrument and troubleshoot it with OpenTelemetry, Application Insights/Azure Monitor, and Grafana. The post focuses on KV-cache behavior under agentic load, dashboard pitfalls, and a vLLM KV-cache FP8 setting that improves concurrency and time-to-first-token.

GisliGudmundsson shares a reference pattern for running Azure AI Foundry workloads inside a secure Azure landing zone, using private endpoints and private DNS zones, disabling local auth in favor of managed identity, and placing Azure API Management in front to enforce token limits and emit token usage metrics.

Tamir Dresher explains where Durable Task (and Durable Task Scheduler) actually helps in production: coordinating long-running workflows that need durable waits, replay, and recovery. Using five concrete scenarios—including Azure tenant provisioning and an AI-assisted incident workflow—he shows what durability solves and what your app still must own.

This week's Weekly AI Roundup focuses on what it takes to run agents like real software: governed tool access, enforceable runtime boundaries, and measurable cost and quality. GitHub Copilot expanded practical model choice (including Claude Haiku 5.5), brought MXC-powered local sandboxing to GA, and added tighter enterprise controls for JetBrains, code review billing, and usage metrics. On Azure, the story moved toward durable execution and production integrations with Dynamic Workflows, Python agent bindings, and managed connectors, backed by tracing patterns in Foundry and OpenTelemetry-based TokenOps. Security and evaluation rounded out the week with a purpose-built secret detection model, new lessons from AI-assisted vulnerability research, and shared benchmarks and playbooks (ReviewBench and the Agent Experience approach) to make agent behavior testable and improvable.

Rama Bhimanadhuni explains how Microsoft is working with the Open Compute Project (OCP) and major hardware vendors to standardize reliability and manageability for large, multi-vendor AI clusters, focusing on RAS signals, diagnostics, telemetry, and safer firmware update workflows to keep fleet capacity and availability high.

Allison announces updates to GitHub Copilot for JetBrains, including enterprise-managed default agent models, a new “Fix” action from diagnostics that opens inline chat, and controls for MCP server startup. The release also improves chat navigation and account UX, boosts reliability, and drops support for JetBrains IDE 2025.1.

GitHub hosts a two-hour live-coding session with Wes Bos, Burke Holland, and Pierce Boggan showing how GitHub Copilot can take a 3D foldable phone simulator from scratch to running functional web apps, including parallel agent workflows, model choices, and practical debugging in the browser.

Wuyi Weng shares a practical, numbers-driven guide to self-hosting the Qwen3.8-27B model on an Azure H100 Spot VM, including the exact vLLM serving flags, a Docker Compose deployment, and an end-to-end observability setup with OpenTelemetry, Application Insights, and Grafana to decide when Spot hosting breaks even.

Allison summarizes the October 5 weekly GitHub Copilot releases, covering a new Claude Haiku 5.5 model option, generally available local sandboxing controls for agent sessions, Copilot app account separation for licensing vs repos, Copilot CLI local model discovery via Ollama, and new VS Code agent-session management features.